Observability

Why observability matters

In enterprise settings, AI must be not only usable but governable:

• Compliance audits — every judgment needs a paper trail
• Incident triage — why did the Agent decide this?
• Cost governance — token spend, model calls, external API usage
• Quality traceability — every step toward a conclusion is replayable

Four observability layers

Session replay

Every session can be fully replayed:

The replay panel includes:

• Timestamp for each step
• Each model call (prompt + completion)
• Each tool call (arguments + return value)
• Token usage and cost per step

This enables minute-level incident reproduction and compliance auditing.

Ledger: disciplined step execution

Ledger is the Orchestrator’s core component—it records incrementally:

• step_status tracking (pending / running / success / failed)
• Confidence for each matched item
• Low-confidence semantic backtracking for multi-turn validation
• Discipline checks and replan fuse conditions

Agents cannot skip Ledger writes—it is write-then-act, not “log when done.”

Distributed tracing

Built on OpenTelemetry:

• Trace ID spans the full request path
• Calls across agents, sub-agents, and external APIs are linked
• Integrates with standard APM systems (Jaeger, Tempo, DataDog, etc.)

Cost visibility

Aggregate by:

• Tenant / user / session / skill
• Model / time window
• Success / failure

Useful for model cost optimization and budget control.

Related docs

• Hooks — low-level entry point for observability hooks
• Monitoring & operations
• Security & compliance